ViViT fall detection and action recognition
Takashi Higashi, Ryuto Ishibashi, Lin Meng · 2024
In Japan, the medical and nursing care industry is facing a serious shortage of medical and nursing care workers due to a decrease in the number of medical and nursing care workers as the population ages and the workforce shrinks due to the declining birthrate and aging population. In addition, falls are an indicator of certain medical conditions and health problems, and early detection has the potential to enhance medical intervention. Fall detection technology could contribute to improving the safety of the elderly and the efficiency of caregiving. Therefore, building a system for fall detection is important for society and can be applied and contributed to various fields. In this research, a video recognition AI using deep learning will be used to automate fall detection. Specifically, the system recognizes falls and other human behaviors using video recognition with ViViT for fall detection and behavior recognition. Furthermore, based on these recognition results and those of conventional video recognition AI, its accuracy and processing speed will be evaluated. The experimental results showed the effectiveness of video recognition using ViViT, with fall detection showing Accuracy:98.88%,Recall:96.67%,Precision:92.36%,F-score:95.72%,Throughput:5782.80qps. The 10-class action recognition was Accuracy:74.73%. Future work includes updating the dataset and multimodal to improve accuracy and token merging to improve computational efficiency.